Human inspired fall prediction method for humanoid robots

Abstract

Humanoid robots are anticipated to work like humans in unstructured environments, and in such cases, falling over is inevitable due to the inherent postural instabilities and external disturbances arising from the environments. Since falling over may annihilate both the robot and its surroundings, we introduce in this work a generic method to predict the falling over of humanoid robots in a reliable, robust, and agile manner across various terrains, and also amidst arbitrary disturbances. The aforementioned characteristics have been strived to attain by proposing a prediction principle inspired by the human balance sensory systems. Accordingly, the fusion of multiple sensors such as inertial measurement unit and gyroscope (IMU), foot pressure sensor (FPS), joint encoders, and stereo vision sensor, which are equivalent to the human’s vestibular, proprioception, and vision systems are considered. For the prediction process, we first define a set of feature-based fall indicator variables (FIVs) from the different sensors, introduce prediction window parameters, and the thresholds for the FIVs are extracted using those parameters for four major disturbance scenarios. Further, an online threshold interpolation technique and an impulse adaptive counter limit are proposed to manage more generic disturbances. Finally, an instantaneous integer value is computed for each FIVs using their respective thresholds and the cumulative sum of them are normalized to predict the fall over by setting a suitable value as the critical limit. To determine the best combination and the usefulness of multiple sensors, the prediction performance is evaluated on four different types of terrains, in three unique combinations: first, each feature individually with their respective FIVs; second, an intuitive performance-based (PF); and finally, Kalman filter based (KF) techniques, which involve the usage of multiple features. For PF and KF techniques, prediction performance evaluations are carried out with and without adding noise to ascertain the influence of sensor noise. Overall, it is reported that KF performed better than PF and individual sensor features under different conditions. Also, the method’s ability to predict fall overs during the robot’s simple dynamic motion is also tested and verified through simulations. Experimental verification of the proposed prediction method on flat and uneven terrains is carried out with the WALK-MAN humanoid robot.